Convolutional LSTM Forecasting of Methane Concentrations Using Copernicus CAMS Data for Relative Priority Screening of Regions in Java Island

Abstract

Methane (CH4) is a potent greenhouse gas, and next-day forecasts of its concentration field can support the allocation of additional monitoring effort. This study forecasts daily column-mean CH4 over Java Island with a Convolutional Long Short-Term Memory (ConvLSTM) network using Copernicus Atmosphere Monitoring Service (CAMS) fields, which are model forecasts and not satellite retrievals. Carbon monoxide (CO), an atmospheric pollutant and indirect climate-forcing agent, is used only as an auxiliary predictor. The data cover 1 March 2024 to 12 July 2026 on a 35 × 96 grid (0.1°, about 11 km) with 2,903,040 cell-days. Models use a seven-day window to predict the next day, a chronological 70/15/15 split and training-only scaling, and are compared in physical units with persistence, climatology, seasonal-naive, ridge regression, gradient boosting and a non-recurrent CNN. The variant selected on the validation block, a residual ConvLSTM with CH4 and CO inputs, reached a test MAE of 7.15 ppb, RMSE of 9.89 ppb and R2 of 0.618 (anomaly R2 0.428), against 7.43 ppb, 10.35 ppb and 0.582 for persistence (MSE skill score 0.087). Given the short record (583 training days), the margin is small: ridge regression performed comparably (MAE 7.02 ppb), and a rolling-origin test showed no gain over persistence in the earliest fold. Area-weighted aggregation to 119 districts and cities gave relative priority classes that largely reproduce the persistent spatial pattern, with the Jakarta metropolitan area ranked highest. The classes are quartile ranks for screening, not validated risk levels or source attributions.